Just an FYI:
I was working on the nucleii dataset from datasciencebowl2018.
I innocently chose transforms_top_down as the augmentation since it seemed natural.
the results were disastrous. the masks were no longer matching images.
I tried re-constructed transforms_top_down (for the nucleii dataset from datasciencebowl2018) as:
aug_tfms = [RandomRotate(10, tfm_y=TfmType.CLASS),
That works just fine.
So don’t use transforms_top_down as is and you can not specify it as: